easyvolcap: switching back to ujson due to api mismatch & adding easy…

…volcap_to_nerf camera parameter conversion

easyvolcap/utils/console_utils.py

@@ -29,7 +29,7 @@ class Colors:
 essential_packages = [
     'pdbr', # will also install rich
     'tqdm',
-    'orjson',
+    'ujson',
     'ruamel.yaml',
 ]
 
@@ -57,7 +57,7 @@ class Colors:
 import warnings
 import readline  # if you need to update stuff from input
 import numpy as np
-import orjson as json
+import ujson as json
 from bdb import BdbQuit
 from pdbr import RichPdb
 from ruamel.yaml import YAML

requirements.txt

@@ -40,7 +40,7 @@ ipdb
 pdbr
 ninja
 lpips
-orjson
+ujson
 pandas
 
 # for unwrapping to get StVK properly

scripts/instant-ngp/easymocap_to_colmap.py

@@ -3,7 +3,7 @@
 import argparse
 import collections
 import numpy as np
-import orjson as json
+import ujson as json
 
 from tqdm import tqdm
 from os.path import join

scripts/instant-ngp/merge_optimized_cameras.py

@@ -4,7 +4,7 @@
 import msgpack
 import argparse
 import numpy as np
-import orjson as json
+import ujson as json
 import torch.nn.functional as F
 
 from tqdm import tqdm

scripts/nerf/easyvolcap_to_nerf.py

@@ -0,0 +1,102 @@
+# 将单帧数据转换为Nerf格式
+import os
+import cv2
+import math
+import argparse
+import numpy as np
+from glob import glob
+
+from easyvolcap.utils.console_utils import *
+from easyvolcap.utils.easy_utils import read_camera, write_camera
+
+
+def convert_K(K, D, H, W):
+    fl_x = K[0, 0]
+    fl_y = K[1, 1]
+    cx = K[0, 2]
+    cy = K[1, 2]
+
+    k1, k2, p1, p2, p3 = D[0]
+
+    angle_x = math.atan(W / (fl_x * 2)) * 2
+    angle_y = math.atan(H / (fl_y * 2)) * 2
+    # fovx = angle_x * 180 / math.pi
+    # fovy = angle_y * 180 / math.pi
+
+    out = {
+        "camera_angle_x": angle_x,
+        "camera_angle_y": angle_y,
+        "fl_x": fl_x,
+        "fl_y": fl_y,
+        "k1": k1,
+        "k2": k2,
+        "p1": p1,
+        "p2": p2,
+        "p3": p3,
+        "cx": cx,
+        "cy": cy,
+        "h": H,
+        'w': W,
+    }
+    return out
+
+
+def convertRT(RT0):
+    RT = np.eye(4)
+    RT[:3] = RT0
+    c2w = np.linalg.inv(RT)
+    c2w[0:3, 1] *= -1
+    c2w[0:3, 2] *= -1  # flip the y and z axis
+    # c2w = c2w[[1, 0, 2, 3], :]  # swap y and z
+    # c2w[2, :] *= -1  # flip whole world upside down
+    return c2w
+
+
+@catch_throw
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data_root', type=str, default='/mnt/data/home/xuzhen/projects/large_gaussian/demo/cameras/00')
+    parser.add_argument('--out_file', type=str, default='/mnt/data/home/xuzhen/projects/large_gaussian/demo/trajectories/path0.json')
+    parser.add_argument('--images_dir', type=str, default='images')
+    parser.add_argument('--img_ext', type=str, default='.png', help='Only used when not using a multi_frame_dataset for contructing file_path')
+    parser.add_argument('--image_file', type=str, default='000000.jpg', help='Only used when --multi_frame_dataset for contructing file_path')
+    parser.add_argument('--multi_frame_dataset', action='store_true')
+    args = parser.parse_args()
+
+    os.makedirs(dirname(args.out_file), exist_ok=True)
+    cameras = read_camera(args.data_root)
+
+    camera_names = list(cameras.keys())
+    annots = {'frames': []}
+
+    K = cameras[camera_names[0]]['K']
+    D = cameras[camera_names[0]]['D']
+    H = cameras[camera_names[0]]['H']
+    W = cameras[camera_names[0]]['W']
+    annots.update(convert_K(K, D, H, W))
+
+    for camera_name in tqdm(camera_names):
+        camera = cameras[camera_name]
+
+        if args.multi_frame_dataset:
+            filename = f'{args.images_dir}/{camera_name}{args.img_ext}'
+        else:
+            filename = f'{args.images_dir}/{camera_name}/{args.image_file}'
+
+        # convert RT
+        c2w = convertRT(camera['RT'])
+        info = {
+            'file_path': filename,
+            'transform_matrix': c2w.tolist(),
+        }
+        info.update(convert_K(camera['K'], camera['D'], camera['H'], camera['W']))
+        annots['frames'].append(info)
+
+    with open(args.out_file, 'w') as f:
+        json.dump(annots, f, indent=4)
+
+    log(yellow(f'Saved converted camera path to: {blue(args.out_file)}'))
+
+
+if __name__ == '__main__':
+    main()

scripts/nerf/nerf_to_easyvolcap.py

@@ -15,6 +15,7 @@ def main():
     parser.add_argument('--transforms_file', type=str, default='transforms_train.json')
     parser.add_argument('--intri_file', type=str, default='cameras_train/00/intri.yml')
     parser.add_argument('--extri_file', type=str, default='cameras_train/00/extri.yml')
+
     parser.add_argument('--images_dir', type=str, default='images_train/00')
     parser.add_argument('--no_organize_images', action='store_false', dest='organize_images')
     parser.add_argument('--no_convert_test', action='store_false', dest='convert_test')
@@ -65,7 +66,7 @@ def main():
     )
     log(yellow(f'Converted cameras saved to {blue(join(args.volcap_root, f"{{{args.intri_file},{args.extri_file}}}"))}'))
 
-    if args.organize_images:
+    if args.organize_images and len(transforms.frames) and exists(transforms.frames[0].file_path):
         run(f'python scripts/nerf/organize_images.py --nerf_root {args.nerf_root} --volcap_root {args.volcap_root} --transforms_file {args.transforms_file} --images_dir {args.images_dir}')
 
     if args.transforms_file.endswith('_train.json') and \

scripts/torchxpbd/register_smplh_semantically.py

@@ -4,7 +4,7 @@
 import torch
 import argparse
 import numpy as np
-import orjson as json
+import ujson as json
 
 from os.path import join
 from torch.optim import Adam